↑↓ Navigateenter Openesc Close
Dashboard
MODEL IDtopazlabs:wonder@3.5
live

Topaz Labs Wonder 3.5

Topaz Labs
by Topaz Labs

Wonder 3.5 is Topaz Labs' updated generative image enhancement and upscaling model for recovering detail and realism from both high-quality and degraded source images. It builds on Wonder 3 with finer detail generation, fewer repetitive patterns, improved handling of visual noise, and better precision on structured content such as tables, charts, and text-heavy graphics. It is well suited to photo restoration, compressed-image cleanup, creative enlargement, and enhancement workflows that need stronger texture recovery without drifting into overly artificial output.

Topaz Labs Wonder 3.5

Upscaling images with Wonder 3.5

How to upscale images with Topaz Wonder 3.5: the request shape, choosing an upscale factor, the input and output size limits, and what generative upscaling recovers.

Enlarging an image the ordinary way just resamples the pixels you already have. Push a small photo to four times its size in an editor and you get a bigger version of the same blur, because there is no more real detail to show. Wonder 3.5 takes the other route: it generates the detail a plain resize cannot, rebuilding fabric weave, skin texture, foliage, and clean edges as it scales the image up.

A tan pebbled-leather tote handbag on a grey studio background, soft and low resolution
SourceWonder 3.5
A 512 px product thumbnail rebuilt at four times the size. Drag to compare.

Both frames sit at the same size on the page, so the difference is what each one actually holds. Drag the handle onto the leather and the source runs out of detail past a certain zoom, while the upscaled frame keeps resolving into grain and stitching. A 100% crop of the clasp makes it plain:

The row of top-stitching that was a grey smear in the source comes back as separate stitches, and the pebbled grain reappears across the leather. Wonder did not sharpen those details, it reconstructed them, which is why the result reads as a genuinely larger photo rather than an enlarged thumbnail.

The request

Upscaling is one image in and one image out, and it runs synchronously like any image task: the response carries the finished imageURL directly, no polling. Put the source in inputs.image as a URL, a UUID from an earlier task, or a data URI, name the model, and set how far to scale with upscaleFactor.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'topazlabs:wonder@3.5',
  upscaleFactor: 4,
  inputs: {
    image: 'https://im.runware.ai/image/os/a14d18/ws/2/ii/1a2b3c4d-5e6f-4708-9a1b-2c3d4e5f6071.jpg'
  }
})
import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "topazlabs:wonder@3.5",
            "upscaleFactor": 4,
            "inputs": {
                "image": "https://im.runware.ai/image/os/a14d18/ws/2/ii/1a2b3c4d-5e6f-4708-9a1b-2c3d4e5f6071.jpg"
            }
        })


asyncio.run(main())
curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "upscale",
      "taskUUID": "b7c8d9e0-1f2a-4b3c-9d4e-5f6a7b8c9d0e",
      "model": "topazlabs:wonder@3.5",
      "upscaleFactor": 4,
      "inputs": {
        "image": "https://im.runware.ai/image/os/a14d18/ws/2/ii/1a2b3c4d-5e6f-4708-9a1b-2c3d4e5f6071.jpg"
      }
    }
  ]'
runware run topazlabs:wonder@3.5 \
  upscaleFactor=4 \
  inputs.image=https://im.runware.ai/image/os/a14d18/ws/2/ii/1a2b3c4d-5e6f-4708-9a1b-2c3d4e5f6071.jpg
{
  "taskType": "upscale",
  "taskUUID": "b7c8d9e0-1f2a-4b3c-9d4e-5f6a7b8c9d0e",
  "model": "topazlabs:wonder@3.5",
  "upscaleFactor": 4,
  "inputs": {
    "image": "https://im.runware.ai/image/os/a14d18/ws/2/ii/1a2b3c4d-5e6f-4708-9a1b-2c3d4e5f6071.jpg"
  }
}
Response
[
  {
    "taskType": "upscale",
    "taskUUID": "b7c8d9e0-1f2a-4b3c-9d4e-5f6a7b8c9d0e",
    "imageUUID": "8e9f0a1b-2c3d-4e5f-9a6b-7c8d9e0f1a2b",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/8e9f0a1b-2c3d-4e5f-9a6b-7c8d9e0f1a2b.jpg"
  }
]
  • inputs.image is the only required input. It takes a URL, a UUID from a previous generation, or a base64 data URI.
  • upscaleFactor is the scale multiplier, from 2 to 6, and defaults to 2.
  • settings.enhancementStrength controls how much detail Wonder invents, covered in enhancement strength.
  • settings.grain adds film grain to the result, covered in adding film grain.

Choosing the upscale factor

upscaleFactor multiplies both the width and the height, so a factor of 4 turns a 512 × 512 source into a 2048 × 2048 image and a factor of 2 doubles each edge. The right value is set by how far the source has to travel: a 2× pass gently sharpens an image that is already a usable size, while a 4× or higher pass is what rescues a small thumbnail or readies it for print.

The coastline below is the same idea on a wide landscape frame, a soft 512 px source lifted to 2048 px across:

A soft, low-resolution aerial photo of a turquoise Mediterranean bay and ochre cliffs
SourceWonder 3.5
A small travel photo enlarged 4×, foliage and rock detail rebuilt as it scales.

The ceiling is on the output, which cannot exceed 256 megapixels, so the factor you pick caps the input Wonder will accept. That works out to roughly 64 MP of input at 2× (around 8000 × 8000), about 16 MP at 4× (around 4000 × 4000), and about 7 MP at 6×. A source larger than the cap for the factor you chose is rejected rather than clipped.

The input limit moves with the factor. An 8000 × 8000 source is fine at 2× but too large at 4×, because 4× would push the output past 256 MP. If a big source is refused, either lower upscaleFactor or downscale the input before sending it.

What generative upscaling does, and does not

Wonder rebuilds detail by synthesizing it, so it makes an image look sharp and richly textured rather than producing a forensically exact enlargement. On a reasonable source the invented detail tracks what was really there, which is why the handbag stitching and the coastline foliage read as correct. On a badly degraded source it will still confidently invent texture that looks right but may not match the original subject.

That shapes what to feed it. Wonder lifts a soft or small image well, but it cannot restore detail that was fully destroyed, so enhance a photo while it still carries a usable signal rather than after it has been crushed. Faces and fine text are the least forgiving targets, and each has its own approach in restoring photos and text and graphics.

Tips

  1. Match the factor to the source. Use 2× to firm up an image that is already close to the size you need, and 4× or higher only when a small source has real distance to cover. A bigger factor is not automatically better.

  2. Feed it recoverable input. Wonder rebuilds soft and moderately small images well, but it works from whatever signal survives. Upscale before quality is destroyed, not after.

  3. Mind the size ceiling. The output tops out at 256 megapixels, so the input limit tightens as the factor grows. Check your source against the cap for the factor you plan to use.

  4. Upscale at the end. Do your cropping and colour work first, then upscale as the final step so the added detail lands on the finished frame rather than getting resampled again downstream.

  5. Judge the result at 100%. A page-width preview hides both the gains and the artefacts. Open the full image or a crop at native size to see what Wonder actually reconstructed.

  6. Reach for the other controls when you need them. Dial the invented detail up or down with enhancement strength, and restore an organic, photographic texture with film grain.

Ask AI
Send feedback